← Back to Library

How far can Chinese hbm go?

Jordan Schneider delivers a sobering reality check on the semiconductor battlefield, arguing that the bottleneck for China's artificial intelligence ambitions isn't just about logic chips, but the invisible, stacked memory that feeds them. While much of the geopolitical discourse fixates on the latest export bans on logic processors, Schneider reveals that the true choke point lies in High Bandwidth Memory (HBM), a component where China faces a generational gap that cannot be easily bridged by domestic ingenuity alone. This is not a story of inevitable convergence, but of specific, hard engineering walls that export controls have effectively turned into cliffs.

The Memory Wall

Schneider frames the problem with surgical precision: AI operations are often "memory constrained," meaning that powerful compute units sit idle, waiting for data to be fed by slower memory chips. He writes, "HBM was created to address this 'memory wall' by stacking multiple memory chips on top of each other to boost memory bandwidth." This distinction is crucial for understanding why simply making more chips isn't enough; the architecture requires a specific type of speed and density that China's current champion, CXMT, struggles to replicate.

How far can Chinese hbm go?

The author paints a stark picture of the competitive landscape. "The three memory giants of SK Hynix, Samsung, and Micron continue to be more than two generations ahead of CXMT's HBM2," Schneider notes. This gap is not merely a matter of time; it is a matter of physics and equipment access. He points out that even if China manages to achieve the industry's current HBM3E or HBM4 standards, "competitive AI chips will likely have meteored beyond contemporary standards to handle workloads unimaginable today." This suggests a moving target that may render China's catch-up efforts obsolete before they even reach the finish line.

Simply put, if you care about the AI race and AI chips, then you must care about HBM.

The Lithography Trap

The core of Schneider's argument rests on the technical specifics of DRAM manufacturing, where he identifies lithography as the primary adversary. He explains that while the industry has moved from numerical nodes to Greek letters (1α, 1β, 1γ), the physical requirements for shrinking these cells are becoming impossible without Extreme Ultraviolet (EUV) lithography. "Without EUV, advanced nodes will either be impossible to make or of terrible yield," he asserts, citing estimates that EUV saves 3-5% yield while cutting process steps by 20-30%.

Schneider draws a parallel to the logic chip sector, noting that while SMIC has managed to stretch Deep Ultraviolet (DUV) tools to create 7 nm chips for Huawei, the memory industry faces different constraints. "Micron has used techniques like self-aligned quadruple patterning (SAQP) to continue to use DUV up until its 1β node," he writes, suggesting CXMT might do the same. However, the ceiling is clear: "1γ and beyond will become extremely difficult without access to the export-controlled EUV equipment." This analysis effectively reframes the narrative from a race of innovation to a race against a hard technological limit imposed by the lack of specific tools.

Critics might argue that China's history of rapid industrial scaling suggests they could find workarounds or that domestic equipment manufacturers are closer to parity than Western estimates allow. However, Schneider's granular breakdown of the bitline contact and storage node contact processes suggests that the precision required for these layers is simply beyond the reach of current non-EUV methods without catastrophic yield losses.

The Packaging and Base Die Dilemma

Beyond the memory dies themselves, Schneider highlights the often-overlooked complexity of the base die and the packaging that binds the stack together. The base die acts as the router for the memory stack, and as generations advance, it requires increasingly sophisticated logic nodes. "Around the HBM4 generation, though, memory makers are compelled to use more expensive logic nodes to handle the workload," he explains. This forces a dependency on advanced logic manufacturing, a sector where the gap between China and the West is even wider.

He illustrates the performance penalty of cutting corners here: "While Micron's product meets the JEDEC minimum of 8 Gbps per pin and goes to 9 Gbps, SK Hynix and Samsung have been able to reach 10 Gbps per pin and beyond via logic node base dies." For CXMT, this means that even if they can manufacture the memory dies, the base die may limit the entire stack's performance. "For CXMT, this likely means that using 1β DRAM dies for HBM4 will result in a subpar product," Schneider concludes, noting that the path to HBM4E is blocked by the need for advanced logic nodes that require EUV.

Interestingly, the article touches on Through-Silicon Vias (TSVs), the vertical connections essential for stacking. Here, the outlook is more optimistic for China. "Having already achieved likely self-sufficient capabilities in TSV formation, CXMT will not be bottlenecked from this step in HBM manufacturing," Schneider writes. This nuance is vital; it shows that the blockade is not uniform. China can handle the vertical connections, but the horizontal logic and the memory cell shrinking remain the critical failure points.

Without EUV, advanced nodes will either be impossible to make or of terrible yield.

The Human and Economic Stakes

While the article is deeply technical, the implications ripple far beyond the semiconductor fab. The inability to produce advanced HBM effectively locks China out of the most demanding AI training workloads, potentially stalling their broader economic and strategic ambitions in artificial intelligence. Schneider's analysis suggests that the "memory wall" is not just a technical hurdle but a strategic moat. The administration's export controls, by targeting the specific equipment needed for EUV and advanced etching, have successfully identified and exploited the weakest link in China's supply chain.

Bottom Line

Schneider's most compelling contribution is his dismantling of the idea that China can simply "out-innovate" its way out of equipment shortages; the physics of advanced memory manufacturing simply do not allow for it without the specific tools the West has withheld. The argument's greatest strength is its granular focus on the base die and lithography constraints, which are often glossed over in favor of broader logic chip narratives. However, the analysis assumes that domestic Chinese equipment will not make a sudden, unforeseen leap in capability, a risk that remains the only variable capable of altering this trajectory.

The verdict is clear: the path to advanced AI in China is not just blocked by a wall, but by a series of increasingly narrow gates that the current administration has effectively locked. The world should watch not for a breakthrough in Chinese memory, but for how long the rest of the industry can maintain its lead while China attempts to build a ladder out of a pit it cannot dig its way out of.

Deep Dives

Explore these related deep dives:

  • High Bandwidth Memory

    The article centers on HBM technology but assumes reader familiarity with its technical architecture. Wikipedia's article explains the stacking technology, TSV interconnects, and generational differences (HBM, HBM2, HBM3) that would deepen understanding of why this is such a difficult manufacturing challenge.

  • Extreme ultraviolet lithography

    The article identifies EUV as the critical chokepoint preventing China from advancing beyond certain DRAM nodes. Understanding the physics of EUV, why ASML has a monopoly, and the decades of development required would illuminate why this export control is so effective.

  • Dynamic random-access memory

    The article discusses DRAM manufacturing nodes, capacitor structures, and wordlines/bitlines without explaining how DRAM fundamentally works. Understanding the one-transistor one-capacitor cell structure and refresh requirements would help readers grasp why scaling is so challenging.

Sources

How far can Chinese hbm go?

by Jordan Schneider · ChinaTalk · Read full article

This December, we’re teaming up with GiveDirectly to send cash to 800 impoverished families in the Bikara region of Rwanda. Studies show that direct cash transfers have a multiplier effect of 2.5x in local economies and reduce infant mortality rates by 48%. Your donation is also tax-deductible in the United States. The link to give is here, and the deadline for donations is midnight on December 31st. Please consider donating if you can!

Aqib Zakaria is a researcher focused on semiconductors, AI, China, and Taiwan. He holds a Master’s degree in Regional Studies — East Asia from Harvard and was recently a summer fellow at the Centre for the Governance of AI (GovAI).

High-bandwidth memory, or HBM, remains the key bottleneck for China to catch up in manufacturing advanced AI chips. As Moore’s Law has more or less held steady, logic nodes have continuously progressed.

However, the rate of memory chip progression has been slow compared to logic chips. Thus, AI operations are often “memory constrained,” meaning that compute is sitting idle waiting for the memory chip to feed it data on which to perform operations. HBM was created to address this “memory wall” by stacking multiple memory chips on top of each other to boost memory bandwidth. As AI chips continue to get better, HBM remains a critical component for scaling. Simply put, if you care about the AI race and AI chips, then you must care about HBM.

Although China’s memory champion CXMT has been closing the HBM gap, the three memory giants of SK Hynix, Samsung, and Micron continue to be more than two generations ahead of CXMT’s HBM2. Assuming export controls hold steady, China’s HBM advances will continue to be stymied by a lack of advanced equipment.

For perspective, achieving the industry’s current HBM3E and HBM4 would be a tremendous achievement for China. As of November 2025, the most advanced AI chips in use use HBM3E. H100s, B100s, and other leading GPUs tap into HBM3E for memory, while Nvidia’s upcoming Rubin GPUs will use HBM4. If CXMT can achieve HBM4 quickly, then they will be able to crack a key part of making advanced GPUs. However, even if they are able to make HBM4 several years down the line, competitive AI chips will likely have meteored beyond contemporary standards to handle workloads unimaginable today.

Ray Wang’s piece earlier this year in ChinaTalk mapping CXMT alongside other ...